Unverified50% confidenceFactExact time
The ByteDance team's control experiments trained from scratch using the Qwen3-0.6B architecture show that all models using chunked compression exhibit periodic accuracy fluctuations, with the fluctuation period strictly equal to the compression step; the fluctuations still exist after removing RoPE or switching to evenly distributed gating weights; and the uncompressed full-attention baseline is completely normal.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedComparing a specifically trained 0.8B model against an untuned stock Qwen3.5-2B base model presents an issue of unequal comparison baselines.71% similarUnverifiedStopping condition design typically faces four categories of trade-off mechanisms: hard resource limits, oscillation detection, performance thresholds, and diminishing returns detection — sharing the same origin as the early stopping strategy in neural network training.68% similarUnverifiedAs long as chunked compression is done by fixed length, the phase sensitivity problem cannot be avoided; this is a structural flaw of the scheme itself and cannot be eradicated by tuning parameters, changing positional encoding or post-training.67% similarUnverifiedThrough post-training methods, Swift-Qwen3.8-27B reduces thinking tokens by 58.3% and improves inference speed by 1.95×, while keeping accuracy loss within 1%.67% similarUnverifiedExperiments show that repeating training on the same data for more than 4 epochs yields diminishing performance gains and can sometimes result in slight performance degradation.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1012264API
curl https://kongchang.com/api/v1/knowledge/claims/1012264MCP
get_claim(id=1012264)